OwnerSunshine530 84db692f07 feat: INT8 dense MoE(torch._int_mm,2D拼接W1_cat/W2_cat,top-k加权折进GEMM2,per-tensor激活量化) ...

dense MoE两个batched GEMM重写成2D GEMM以用A800 int8 tensor core;计算减半。
quant/dequant是真compute本地可见→本地bench即可判生死。默认关,bench --moe-int8。

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-06-20 01:35:55 +08:00

..

tests

perf: _triton_block_meta 消除最后一个host同步(grid用shape派生上界,空block在kernel内mask空跑)

2026-06-19 20:51:37 +08:00

bench.py

feat: INT8 dense MoE(torch._int_mm,2D拼接W1_cat/W2_cat,top-k加权折进GEMM2,per-tensor激活量化)

2026-06-20 01:35:55 +08:00

build_env.sh

fix: build_env.sh 简化为纯净版本（避免 CUDA 预热导致异常）

2026-06-12 21:55:09 +08:00

EXPERIMENTS.md

docs: 收尾 — 最终67.998/记录RepEncoder预计算尝试与结论

2026-06-16 13:18:48 +08:00

infer.py

feat: INT8 dense MoE(torch._int_mm,2D拼接W1_cat/W2_cat,top-k加权折进GEMM2,per-tensor激活量化)

2026-06-20 01:35:55 +08:00

requirements.txt

revert: requirements.txt 还原为原始完整依赖列表

2026-06-12 21:24:22 +08:00

RISKS.md

docs: 潜在风险说明(RepEncoder预计算合规灰区/max_feasign一致性)与合规保底

2026-06-15 20:44:57 +08:00